File auditing method, device and system and medium

Abstract

The invention provides a file auditing method, device and system and a medium, and relates to the technical field of file processing.The method comprises the steps that when a target task chain dynamically generated according to a file auditing intention comprises a sensing component, a verification component, a cognition component and a decision component, the target task chain is determined on the basis of the sensing component; performing structured information extraction on the to-be-audited bill picture in the to-be-audited file to obtain structured bill information and identification confidence of the structured bill information; based on the verification component, performing compliance verification on the structured bill information to obtain verification confidence; based on a cognitive component, performing consistency and compliance reasoning on the structured bill information and reimbursement reason information in the to-be-audited file to obtain cognitive confidence; and based on the decision component, combining the recognition confidence coefficient, the verification confidence coefficient and the cognition confidence coefficient, and performing decision to obtain a final audit result. According to the method, the dynamic task chain drives the multiple agents to carry out multi-modal auditing decision, so that the reliability and the flexibility of file auditing are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of file processing, and in particular to a file auditing method, device, system and medium. BACKGROUND

[0002] With the deepening of enterprise digital transformation, financial automation auditing has become an important means for enterprises to reduce costs and increase efficiency. In the enterprise financial reimbursement scene, a large number of bills and corresponding reimbursement application forms need to be processed. Therefore, how to automatically audit the bills and corresponding reimbursement application forms to ensure the compliance of the reimbursement files is particularly important.

[0003] At present, the conventional auditing method usually adopts a fixed single processing logic for file auditing, such as using optical character recognition technology to extract text information in the bill image, and then using pre-prepared rules or databases to verify the compliance of the structured data such as the amount and date in the bill image based on hard-coded logic conditions. For unstructured text such as the reimbursement reason in the reimbursement application form, a keyword matching method is usually used to detect whether it contains a pre-set violation vocabulary, and finally the verification results of each link are spliced to output an audit conclusion. However, the processing logic of this kind of auditing method is fixed single processing logic, which leads to poor flexibility and reliability in the dynamic and complex reimbursement file auditing scene. SUMMARY

[0004] The present application provides a file auditing method, device, system and medium to solve the defect that the reimbursement file review is performed by the auditing method using a fixed single processing logic in the prior art, which leads to poor flexibility and accuracy, and to improve the reliability of file auditing while improving the flexibility of file auditing in the dynamic and complex reimbursement file auditing scene.

[0005] The present application provides a file auditing method, comprising: generating a target task chain according to a file auditing intention corresponding to target input information; the target input information includes instruction information for configuring a task chain; in the case that the functional components of the target task chain include a perception component, a verification component, a cognition component and a decision component, performing structured information extraction on a to-be-audited bill picture in a to-be-audited file based on the perception component to obtain structured bill information and an identification confidence of the structured bill information; performing compliance verification on the structured bill information based on the verification component to obtain a verification confidence of the structured bill information; performing consistency and compliance reasoning on the structured bill information and reimbursement reason information in the to-be-audited file based on the cognition component to obtain a cognition confidence of the to-be-audited file; Based on the decision component, the identification confidence, the verification confidence and the cognition confidence are combined to obtain an audit result of the file to be audited.

[0006] According to the file audit method provided by the application, the structured invoice information and the reimbursement reason information in the file to be audited are subjected to consistency and compliance reasoning to obtain a cognition confidence of the file to be audited, which comprises: According to the type of the invoice and the industry type to which the seller belongs in the structured invoice information, a plurality of source information associated with the file to be audited is obtained; the plurality of source information comprises enterprise qualification data, historical audit case records, compliance strategy knowledge base and sensitive word library; According to the identification confidence and the verification confidence, target constraint information is generated. The structured invoice information, the reimbursement reason information, the plurality of source information and the target constraint information are input into an inference model to obtain the cognition confidence.

[0007] According to the file audit method provided by the application, the structured invoice information, the reimbursement reason information, the plurality of source information and the target constraint information are input into an inference model to obtain the cognition confidence, which comprises: Each information in the structured invoice information, the reimbursement reason information, the plurality of source information and the target constraint information is input into a feature encoding module corresponding to the data type of each information in the inference model to obtain a feature vector of each information. The feature vectors of each information are input into an attention module in the inference model to obtain semantic similarity between the feature vectors of each information. According to the semantic similarity, the feature vectors of each information are weighted and fused to obtain a fusion feature. The fusion feature is input into an inference module in the inference model to obtain the cognition confidence.

[0008] According to the file audit method provided by the application, the feature vectors of each information are input into an attention module in the inference model to obtain semantic similarity between the feature vectors of each information, which comprises: The feature vectors of each information are input into a self-attention unit in the attention module to obtain local semantic enhanced features of each information. The feature vectors of each information are input into a cross-attention unit in the attention module to obtain cross-modal complementary features of each information. The local semantic enhancement features and the cross-modal complementary features are input into the multi-head attention unit in the attention module to obtain the semantic similarity between the feature vectors of each piece of information.

[0009] According to a document review method provided by the present invention, the step of inputting the fused features into the reasoning module of the reasoning model to obtain the cognitive confidence includes: The fused features are input into the multi-task reasoning unit in the reasoning module to obtain multi-dimensional reasoning results; the multi-dimensional reasoning results include multiple items from consistency verification results, compliance verification results, sensitive content identification results, and contextual reasoning results; The multi-dimensional reasoning results are input into the result aggregation unit in the reasoning module to obtain the cognitive confidence. The consistency verification result includes the degree of content matching between the structured invoice information and the reimbursement reason information, and the degree of qualification matching between the structured invoice information and the enterprise qualification data; the compliance verification result includes the degree of conformity between the reimbursement reason information and each preset compliance strategy data in the compliance strategy knowledge base; the sensitive content identification result includes the degree of matching between the reimbursement reason information and each preset sensitive word in the sensitive word library; the context reasoning result includes the logical reasonableness score of the reimbursement reason information in the current business context obtained by analogy reasoning through the historical audit case records.

[0010] According to a document review method provided by the present invention, the step of performing compliance verification on the structured invoice information to obtain the verification confidence level of the structured invoice information includes: Based on the bill type in the structured bill information, determine the compliance strategy knowledge base and third-party data; Based on the compliance strategy knowledge base and the third-party data, the compliance of the structured invoice information is verified to obtain a first verification result; Based on the historical reimbursement record database, the number of duplicate reimbursements and the frequency of abnormalities in the structured invoice information are verified to obtain a second verification result; Based on a preset quantization mapping strategy, the first verification result and the second verification result are comprehensively scored to obtain the verification confidence level.

[0011] According to a document review method provided by the present invention, the step of generating a target task chain based on the document review intent corresponding to the target input information includes: Based on the document review intent, obtain the task objective, target component type, and logical control conditions; Based on the target component type, retrieve multiple functional components from the prefabricated component library; Based on the stated task objectives, determine the operating parameters of each of the stated functional components; Based on the aforementioned logical control conditions, the data flow paths and execution dependencies between the functional components are determined. Based on the data flow path, the execution dependencies, and the running parameters, multiple functional components are encapsulated to generate the target task chain.

[0012] According to a document review method provided by the present invention, the step of combining the identification confidence level, the verification confidence level, and the cognitive confidence level to determine the review result of the document to be reviewed includes: The recognition confidence, verification confidence, and cognition confidence are weighted and fused to obtain a comprehensive confidence score. If the overall confidence level is greater than or equal to the first threshold, the review result is determined to be a successful review result. If the overall confidence level is less than the first threshold and greater than the second threshold, the review result is determined to be the result that triggers manual review. If the overall confidence level is less than the second threshold, the review result is determined to be a rejection.

[0013] The present invention also provides a document review device, comprising: The generation unit is used to generate a target task chain based on the document review intent corresponding to the target input information; the target input information includes instruction information for configuring the task chain. The perception unit is used to extract structured information from the image of the invoice to be reviewed in the document to be reviewed based on the perception component, when the functional components of the target task chain include a perception component, a verification component, a cognition component and a decision component, to obtain structured invoice information and the recognition confidence of the structured invoice information. The verification unit is used to perform compliance verification on the structured invoice information based on the verification component, and obtain the verification confidence level of the structured invoice information; A cognitive unit is used to perform consistency and compliance reasoning on the structured invoice information and the reimbursement reason information in the document to be audited based on the cognitive component, and to obtain the cognitive confidence level of the document to be audited. The decision unit is used to determine the audit result of the document to be audited based on the decision component, in conjunction with the identification confidence, the verification confidence, and the cognitive confidence.

[0014] The present invention also provides a document review system, including a user interaction module, a task chain engine, a processor, a database, and a model optimizer; The task chain engine includes a memory, an execution engine, and a computer program stored in the memory and running on the processor. When the execution engine executes the computer program, it implements the document auditing method as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the document auditing method as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the document review methods described above.

[0017] The document review method, apparatus, system, and medium provided by this invention dynamically generate target task chains based on indications in the target input information, breaking the rigid limitations of traditional hard-coded processes and enabling flexible reconstruction of review logic to effectively adapt to dynamic business changes. Simultaneously, by combining perception, verification, and cognition components in the task chain for multi-dimensional in-depth analysis, it fills the semantic understanding gap in traditional keyword matching. Finally, the decision component, in conjunction with the recognition confidence, verification confidence, and cognition confidence outputs of multiple components, conducts a comprehensive judgment, effectively avoiding the one-sidedness of single logical judgments. This significantly improves the flexibility of the review process while greatly enhancing the reliability of identifying hidden risks in complex reimbursement scenarios. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts illustrating the document review method provided by this invention.

[0020] Figure 2 This is a schematic diagram of the document review system provided by the present invention.

[0021] Figure 3 This is a schematic diagram of the task chain construction process provided by the present invention.

[0022] Figure 4 This is a schematic diagram of the structure of the multi-agent system provided by the present invention.

[0023] Figure 5 This is a schematic diagram of the cognitive confidence acquisition process provided by the present invention.

[0024] Figure 6 This is the second flowchart of the document review method provided by the present invention.

[0025] Figure 7 This is a schematic diagram of the document review device provided by the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] All actions involving the acquisition of signal information or data in this application are carried out in accordance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0028] Currently, conventional review methods typically employ a fixed, single processing logic for document review. For example, some technologies use Optical Character Recognition (OCR) to extract text information from invoice images. Then, based on preset rules and hard-coded logical conditions, they perform compliance checks on the structured data such as amounts and dates extracted from the invoice images. Additionally, some technologies build expert knowledge bases to perform integrity checks and logical conflict detection on the structured data extracted from invoice images, extract text information from the reimbursement reason, and simultaneously identify the authenticity of the invoice images. Finally, they combine the verification results from each stage to output the review conclusion.

[0029] However, the above-mentioned review method has a fixed and singular processing logic, which still faces many challenges in practical applications: First, the system is rigid and has high maintenance costs. Hard-coded linear workflow approval systems write approval rules directly into the program code. When the company's reimbursement rules change or new approval dimensions need to be added, it is necessary to rely on software developers to modify, test, and release the code. Business personnel are bound to the IT development cycle, resulting in long response times, such as 1-2 weeks. In addition, the transformation costs are high, and there is a lack of flexibility, making it difficult to quickly adapt to dynamically changing business needs.

[0030] Secondly, the technical components are isolated and lack coordination. Existing solutions are mostly isolated technology stacks, with expert databases and OCR modules typically only using simple feature concatenation, lacking organic collaboration. Especially for scenarios requiring complex semantic understanding (such as the rationality analysis of reimbursement reasons) and external knowledge verification (such as verifying the business qualifications of sales parties), the capabilities of a single technical means are limited, and dynamic weighted fusion of multi-source information is not achieved, making it impossible to form a comprehensive judgment and resulting in poor reliability.

[0031] Secondly, the level of intelligence in reviewing reimbursement reasons is low. Traditional technologies mainly rely on keyword matching, which cannot understand the contextual semantics of the reason, cannot judge the rationality of the consumption scenario, and cannot cross-compare information such as the amount and date mentioned in the reason with the data on the invoice itself, resulting in weak risk control capabilities and poor reliability.

[0032] In summary, the existing review methods suffer from poor flexibility and reliability in the context of reviewing complex and dynamically changing expense reimbursement documents.

[0033] To address the aforementioned issues, this invention provides a document review method that achieves flexible configuration and deep intelligent reasoning of the review process through a dynamically reconfigurable task chain and the collaborative work and decision fusion of multiple agents, thereby achieving a dual improvement in the reliability and flexibility of document review.

[0034] Figure 1 This is one of the flowcharts illustrating the document review method provided by this invention.

[0035] It should be noted that, as Figure 1 As shown, the document review method provided in this application can be applied to a document review system, which may be a server, cloud platform, or high-performance computing device, etc. This embodiment does not specifically limit it.

[0036] To better understand the method of this embodiment, the system architecture for implementing the method will first be described.

[0037] Figure 2 This is a schematic diagram of the document review system provided by the present invention.

[0038] like Figure 2As shown, the document review system adopts a layered and modular design architecture, thereby achieving fully automated processing from uploading reimbursement documents to outputting compliance review results. Specifically, the system is divided into five main layers: the user interaction layer (also known as the user interaction module), the task chain engine layer (also known as the task chain engine), the multimodal processing layer (also known as the processor), the rule and knowledge base layer (also known as the knowledge base), and the model training and optimization layer (also known as the model optimizer). Each layer communicates loosely through standardized interfaces, such as the Representational StateTransfer Application Programming Interface (REST API) and message intermediaries, supporting high concurrency and high availability deployment. This forms a complete closed-loop intelligent review process including user input, process orchestration, multimodal understanding, rule comparison, intelligent decision-making, and result feedback, achieving intelligent, automated, interpretable, adaptive, and reliable compliance review of reimbursement documents.

[0039] The user interaction layer serves as the front-end entry point and display interface of the document review system. It receives user-uploaded invoices and reimbursement details, and displays the review results. The user interaction layer mainly includes a user interface, an invoice upload module, an API gateway, and a result display module, among other components. This embodiment does not specifically limit these components. The user interface provides a graphical user interface, allowing users to upload invoice images and fill in supplementary information such as the reimbursement details. The invoice upload module receives user-submitted image or Portable Document Format (PDF) invoice files and performs preliminary verification and preprocessing. The API gateway, as the unified access point for the system's external services, is responsible for request routing, authentication, traffic control, and security protection. The result display module presents the final review results to the user in a structured format, including compliance status, risk warnings, and suggested modifications.

[0040] The task chain engine layer enables flexible configuration and dynamic execution of the review process, possessing high scalability and visual management capabilities. This layer primarily comprises a task chain designer (also known as the designer), a component library, a task chain execution engine, and a status monitoring module. The task chain designer allows users to customize review task flows through drag-and-drop, supporting node addition, deletion, modification, and querying, thus improving process configuration efficiency. The component library encapsulates various reusable functional components, such as perception, verification, cognition, decision-making, and rule loading components, for use in process design. The task chain execution engine parses and executes the task flow graph generated by the designer, scheduling each functional component sequentially to ensure orderly process operation. The status monitoring module tracks the task execution progress, resource consumption, and anomalies of each functional component in real time, supporting logging and alarm mechanisms.

[0041] The multimodal processing layer focuses on the fusion processing of multimodal data to achieve cross-modal reasoning, including optical character recognition module, multi-source information fusion module, attention module and large language model reasoning module, etc. This embodiment does not specifically limit these.

[0042] The rules and knowledge base layer provides the static rules and dynamic knowledge support required for auditing, ensuring the consistency and legality of audit standards. It mainly includes an audit rule base, a sensitive word database, a compliance policy database, and external application programming interfaces (APIs). The audit rule base stores various reimbursement audit rules established internally by the company, such as amount limits and invoice type requirements. The sensitive word database contains keywords that may involve violations, such as fake invoices and fictitious transactions, used for quickly screening high-risk invoices. The compliance policy database integrates relevant policy documents such as tax regulations and business administration systems, supporting automatic matching and citation of policy basis. The external application programming interfaces (APIs) connect to third-party government platforms, such as business registration, tax, and credit platforms, to obtain data such as company registration information, invoice authenticity verification, and tax credit ratings, enhancing the credibility of the audit.

[0043] The model training and optimization layer is dedicated to achieving iterative evolution of the model through online learning and reinforcement feedback mechanisms. Specifically, it includes, but is not limited to, components such as the experience pool replay module, supervised training module, reinforcement learning module, and large model fine-tuning module, so as to realize dynamic training and optimization of the models within each module configured in the multimodal processing layer.

[0044] It should be noted that the execution entity of the method provided in this application can be a task chain engine.

[0045] like Figure 1 As shown, the method includes steps 110, 120, 130, 140 and 150.

[0046] Step 110: Generate a target task chain based on the document review intent corresponding to the target input information; the target input information includes instruction information for configuring the task chain.

[0047] Figure 3 This is a schematic diagram of the task chain construction process provided by the present invention.

[0048] like Figure 3 As shown, after receiving the target input information, the target input information can be parsed to obtain the corresponding document review intent.

[0049] like Figure 3 As shown, in this embodiment, the target input information refers to the information submitted by the user through the user configuration interface of the user interaction layer, which is used to trigger and instruct the dynamic configuration of the task chain. This user configuration interface provides a concise interactive entry point. The target input information is the direct carrier of the user's intent. It can be configuration information formed by the user selecting or dragging on the graphical interface, or it can be a natural language command entered by the user through the natural language configuration input interface. This natural language configuration input interface allows the user to enter the target input information for requesting changes to the audit strategy (i.e., dynamic configuration of the task chain) in natural language form. For example, the target input information can be natural language commands such as "add a budget amount check for catering invoices" or "add a corporate credit blacklist verification step," so that even non-technical personnel can trigger adjustments to complex audit processes through simple text commands.

[0050] Document review intent refers to the specific review strategy and logical requirements determined after semantic parsing or code decoding of the target input information. It is used to indicate information such as task objectives, required component types, and logical control conditions. For example, when the target input information is a natural language instruction input by a user, a natural language parser can be invoked to perform semantic understanding and intent recognition on the natural language instruction through a pre-trained language model, thereby extracting the document review intent. The pre-trained language model here refers to existing language models that rely on massive corpus data for pre-training, such as the Bidirectional Encoder Representations from Transformers (BERT) model, generative pre-trained Transformers, text-to-text Transformers, and large-scale pre-trained language models deployed in various interactive systems for human-computer interaction. This embodiment does not specifically limit this type of pre-trained language model.

[0051] After parsing the document review intent, a corresponding task chain, i.e., the target task chain, can be dynamically generated based on that intent. This process can involve automatically inferring component combinations, parameter configurations, and connection relationships from the document review intent to build a completely new task chain from scratch, supporting more complex dynamic evolution and intelligent completion; or, leveraging the powerful text generation capabilities of existing large-scale pre-trained language models, directly generating a configuration file describing the task chain structure based on the document review intent to generate the target task chain, thereby achieving a direct transformation from abstract intent to specific execution logic. This embodiment does not specifically limit this approach.

[0052] For example, in one possible embodiment, generating the target task chain based on the document review intent corresponding to the target input information includes: Based on the document review intent, obtain the task objective, target component type, and logical control conditions; Based on the target component type, retrieve multiple functional components from the prefabricated component library; Based on the stated task objectives, determine the operating parameters of each of the stated functional components; Based on the aforementioned logical control conditions, the data flow paths and execution dependencies between the functional components are determined. Based on the data flow path, the execution dependencies, and the running parameters, multiple functional components are encapsulated to generate the target task chain.

[0053] like Figure 3 As shown, after obtaining the semantic structure of the document review intent output by the natural language parser, the task chain designer can automatically configure the task objectives, target component types, and logical control conditions in the task chain based on the semantic structure of the document review intent, so as to dynamically generate the corresponding control flow without manual intervention. For example, if the document review intent is "add budget limit checks to restaurant invoices", the task objective obtained through semantic analysis is "verify the budget limit of the invoice amount". Based on this task objective, the system matches the metadata in the component library to determine the target component type as a component for budget rule verification used to execute specific verification logic and a component that can call the financial system application programming interface, etc.

[0054] After obtaining the task objective, target component type, and logical control conditions, multiple functional components can be retrieved from the pre-built component library through the task chain designer. This pre-built component library contains standardized, independently encapsulated audit function units, including but not limited to perception components, verification components, cognition components, decision-making components, rule loading components, and enterprise query interface components; this embodiment does not specifically limit these. Each component has clear input / output interfaces and execution contracts, ensuring it can be accurately called and combined by the task chain designer.

[0055] In addition, the conditional branch setting module and parameter configuration module in the task chain designer can automatically configure the running parameters of each functional component according to the task objective. Furthermore, based on the logical control conditions, complex control flows can be dynamically generated to determine the execution dependencies and data flow paths between functional components, thereby achieving transparent data transfer and context sharing between components.

[0056] Subsequently, the version manager encapsulates the above configuration to generate a new task chain version. This task chain structure is displayed through a visual editor, and users can fine-tune the structure via a configuration panel. After confirmation, the configuration is stored in the configuration database via the save configuration module, ultimately forming the target task chain. The functional components in the final target task chain can be scheduled and executed by the workflow executor and task scheduler, managing asynchronous communication, parallel processing, and state transitions between components. Transparent data transfer between components is achieved through data pipelines, supporting multi-format data conversion and context sharing.

[0057] It should be noted that an exception handler can also be attached to the task scheduler to capture, record, and recover error information that occurs during the execution of various functional components, thereby improving system robustness.

[0058] The version manager here supports storing all historical task chain configurations and their version information, and supports one-click rollback, hot deployment and audit tracing to ensure that system changes are controllable and traceable.

[0059] The method provided in this embodiment automatically parses task objectives and logical conditions based on document review intent, and dynamically retrieves components, configures parameters, and plans execution paths accordingly. This achieves end-to-end automated mapping from abstract document review intent to specific task chains. This mechanism deeply decouples review business logic from underlying code, completely eliminating reliance on fixed templates or hard-coded development. Users can reconstruct the review process in real time simply by providing instructions. In other words, users only need to input commands to automatically complete intent recognition, component matching, connection inference, and parameter setting, completely eliminating reliance on code development. This avoids the problem of traditional low-code platforms still requiring manual intervention when facing complex business changes. Even users without programming backgrounds can dynamically adjust and deploy complex review processes through simple text commands to quickly respond to business changes and achieve real-time process updates. This not only significantly reduces operation and maintenance costs and shortens the response cycle from the traditional weekly level to the real-time level, but also supports the dynamic evolution and intelligent completion of task chains, achieving end-to-end automated mapping from abstract business intent to specific task chains, thus improving the flexibility and reliability of review.

[0060] Step 120: If the functional components of the target task chain include a perception component, a verification component, a cognition component, and a decision-making component, then based on the perception component, structured information is extracted from the image of the invoice to be reviewed in the document to be reviewed, to obtain structured invoice information and the recognition confidence of the structured invoice information.

[0061] Optionally, when the functional components of the target task chain include a perception component, a verification component, a cognition component, and a decision component, document review can be achieved through a multi-agent system formed by the perception component (also known as the perception layer), the verification component (also known as the verification layer), the cognition component (also known as the cognition layer), and the decision component (also known as the decision layer) in a way that enables multi-agent collaboration and decision fusion. That is, when the target task chain is started, the automated and highly reliable review of reimbursement documents can be achieved through task linkage and result fusion between multiple functional components.

[0062] Figure 4 This is a schematic diagram of the structure of the multi-agent system provided by the present invention; like Figure 4 As shown, in addition to the perception component, verification component, cognition component, and decision-making component, this multi-agent system may also include an output component (also known as the output layer). The system realizes the collaborative scheduling and result fusion among the agents through the target task chain. It is not simply a stacking of multiple technical means, but rather a well-structured, collaborative, and efficient intelligent auditing system.

[0063] The perception layer is primarily responsible for extracting structured information and its corresponding recognition confidence level from the original document image. The core components of this perception layer include a perceptual agent, a document preprocessing module, an OCR engine, a structured output module, and a quality assessment module.

[0064] The perceptual agent, as the main execution unit of the perception layer, coordinates the preliminary processing and information extraction of the ticket images.

[0065] Accordingly, upon receiving the image of the invoice to be reviewed from the files to be reviewed, the following steps can be executed through a perceptual agent to extract the structured information from the image of the invoice to be reviewed: First, the invoice preprocessing module is invoked to enhance the image of the invoice to be reviewed within the review file. Specifically, to address complex background issues that invoices may have in real-world scenarios, such as tilting, blurring, and uneven lighting, the invoice preprocessing module employs adaptive distortion correction algorithms and illumination equalization techniques to perform geometric correction and color enhancement on the original image. For example, edge detection algorithms are used to locate the four corner coordinates of the invoice and perform perspective transformation to correct distortion, while histogram equalization techniques are used to eliminate the influence of shadows, thereby significantly improving the recognition accuracy of the subsequent OCR engine.

[0066] Subsequently, the preprocessed ticket image is fed into an OCR engine for high-precision recognition. This OCR engine can call an OCR model to perform structured information processing on the preprocessed ticket image. This OCR model can be a single network architecture model formed by an existing single network model, or a joint network architecture model formed by existing multiple network models; this embodiment does not specifically limit it.

[0067] For example, when the OCR model is a joint network architecture model combining the third generation of the document intelligent multimodal pre-trained model (Layout Language Model Version 3, LayoutLMv3) and the Gated Recurrent Unit (GRU) model, the specific steps for extracting structured information include: First, the preprocessed invoice image is input into the LayoutLMv3 model. The LayoutLMv3 model first utilizes its multimodal encoding capability to extract visual features, text features, and layout features from the invoice image simultaneously. Then, it further utilizes its internal spatial attention mechanism to calculate the correlation weights between visual features, text features, and layout features, thereby accurately capturing the layout structure of the invoice in the multimodal feature space and effectively distinguishing key areas in the invoice.

[0068] Then, visual features, text features, and layout structure are input into the GRU network, which performs contextual prediction on each key field to obtain the prediction sequence.

[0069] Subsequently, the output layer of the OCR model decodes the predicted sequence to obtain the content of each key field, such as invoice code, amount, date, seller name, tax ID, and project name. Simultaneously, based on the identified key fields, the invoice type is automatically determined, such as catering, transportation, or office supplies.

[0070] It should be noted that the OCR model here is trained using a model training and optimization layer. Specifically, the training involves: first, loading a labeled dataset specifically designed for expense reimbursement scenarios; then, performing regular supervised training on the initial OCR model based on this labeled dataset to optimize the OCR model's accuracy in recognizing document fields, thus obtaining the final OCR model. This labeled dataset contains multiple sample document images and their structured annotation information.

[0071] Subsequently, the recognition results output by the OCR model are transformed into a standardized data format through the structured output module, forming structured document information. Simultaneously, the quality assessment module quantifies the quality of the recognition results based on a confidence score mechanism. Specifically, during the decoding process, the OCR model generates probability values ​​for each key field. The quality assessment module calculates the recognition confidence score based on these values. This confidence score, as a quantitative indicator of data quality, is passed along with the structured document information to subsequent levels, providing a reliable basis for subsequent decision-making.

[0072] The method provided in this embodiment effectively overcomes the interference of complex shooting environments on ticket recognition by introducing preprocessing techniques such as adaptive distortion correction and illumination equalization into the perception component; and by combining the LayoutLMv3 model and the GRU network OCR engine, the spatial attention mechanism is used to achieve deep fusion of visual and text features, which significantly improves the recognition accuracy of ticket layout structure and key fields.

[0073] Step 130: Based on the verification component, perform compliance verification on the structured invoice information to obtain the verification confidence level of the structured invoice information.

[0074] like Figure 4 As shown, the verification component (also known as the verification layer) focuses on performing compliance verification on the structured ticket information output by the perception component to ensure the legality and compliance of the data. The compliance verification here includes, but is not limited to, hard rule verification and / or external authenticity verification. Specifically, it can be determined based on the running parameters configured for the verification component according to the target task chain.

[0075] The core components of this verification layer include the verification agent, the rule engine, the external application programming interface (API) verification module, the cross-verification module, and the verification result output module.

[0076] The verification agent, as the main execution unit of the verification layer, coordinates the collaborative work of internal rule matching and external interface calls.

[0077] Optionally, after obtaining the structured invoice information, compliance verification can be achieved by scheduling the following steps through a verification agent: First, the structured invoice information is statically validated using a rule engine. Specifically, the rule engine dynamically loads the corresponding audit rules from the compliance policy knowledge base that match the invoice type in the audit rule base, such as rules on amount limits, date reasonableness, and invoice header matching, based on the invoice type in the structured invoice information. These rules are then used to compare the structured invoice information item by item to obtain preliminary rule validation results.

[0078] It should be noted that this rule engine supports dynamic loading and configuration of business rules, and can flexibly adjust logical rules such as amount consistency, date rationality, and invoice header matching to meet the business needs of different enterprises.

[0079] In addition, an external API verification module is used to perform external authenticity checks. This module connects to third-party government platforms such as the business registration and taxation departments through standard interfaces to query third-party data in real time, such as the seller's enterprise qualifications, taxpayer status, and invoice verification information, thereby generating external verification results.

[0080] In addition, the cross-verification module is invoked to perform historical behavior analysis. Based on the historical expense reimbursement record database, this module performs deduplication verification and anomaly frequency analysis on the structured invoice information. For example, it checks whether the same invoice has been submitted repeatedly, or whether the same employee has frequently submitted similar large-amount invoices within a short period of time, thereby generating cross-verification results.

[0081] Subsequently, the verification agent generates a verification confidence level reflecting the degree of data legality and compliance based on at least one of the above rule verification results, external verification results, and cross-verification results, and outputs the verification confidence level to the next level through the verification result output module.

[0082] For example, in one possible embodiment, the compliance verification of the structured invoice information to obtain the verification confidence level of the structured invoice information includes: Based on the bill type in the structured bill information, determine the compliance strategy knowledge base and third-party data; Based on the compliance strategy knowledge base and the third-party data, the compliance of the structured invoice information is verified to obtain a first verification result; Based on the historical reimbursement record database, the number of duplicate reimbursements and the frequency of abnormalities in the structured invoice information are verified to obtain a second verification result; Based on a preset quantization mapping strategy, the first verification result and the second verification result are comprehensively scored to obtain the verification confidence level.

[0083] Optionally, the type of invoice is first identified in order to pinpoint the appropriate subset of the compliance strategy knowledge base and the enterprise qualification data interface to be queried from the rules and knowledge base layer.

[0084] Subsequently, on the one hand, the rules engine is used to compare whether the invoice amount is within the limits specified in the compliance strategy knowledge base; on the other hand, the external API verification module is used to verify whether the seller has the corresponding business qualifications and the authenticity of the invoices, so as to obtain the first verification result.

[0085] Simultaneously, the cross-verification module is invoked to check whether the unique identifier of the invoice already exists and to count the number of times it is repeatedly reimbursed. At the same time, the reimbursement frequency of the person seeking reimbursement within a specific time window is analyzed to determine whether there is an abnormal surge, so as to obtain a second verification result. This second verification result reflects the dynamic risk of reimbursement behavior.

[0086] Finally, a tiered deduction or weighted scoring method is used to quantitatively map the comprehensive score of the first and second verification results. For example, with a base score of 100 points, if the invoice authenticity verification fails in the first verification result (a veto), the score is directly reduced to 0 points; if duplicate reimbursement is found in the second verification result (a hard violation), the score is also reduced to 0 points; if it is only a slightly higher reimbursement frequency (a soft risk), the corresponding weighted score is deducted (e.g., 10 points). The final calculated comprehensive score is the verification confidence level.

[0087] The method provided in this embodiment, by integrating a triple verification mechanism of static rules, external authoritative data, and historical behavior records into the verification component, not only ensures the surface compliance of invoice information, but also effectively identifies hard violation risks such as invoice forgery and duplicate reimbursement through cross-verification of multi-source data, thus ensuring the reliability of invoice recognition.

[0088] Step 140: Based on the cognitive component, perform consistency and compliance reasoning on the structured invoice information and the reimbursement reason information in the document to be audited to obtain the cognitive confidence level of the document to be audited.

[0089] like Figure 4As shown, the cognitive component (also known as the cognitive layer) focuses on understanding and reasoning about unstructured text, and introduces a large language model to achieve in-depth semantic analysis.

[0090] Alternatively, consistent and compliant reasoning can be achieved by invoking the cognitive component to perform the following steps: The structured invoice information and the reimbursement reason information in the documents to be reviewed are directly input into the deep learning model. The deep learning model performs semantic reasoning based on the structured invoice information and the reimbursement reason information to output the corresponding soft semantic reasoning result, i.e., cognitive confidence. Alternatively, based on the structured invoice information and the reimbursement reason information in the documents to be reviewed, contextual environment data and / or pre-confidence constraints are added to the deep learning model to guide the deep learning model to perform semantic reasoning based on the structured invoice information and the reimbursement reason information to output the corresponding soft semantic reasoning result, i.e., cognitive confidence. This embodiment does not specifically limit this.

[0091] The cognitive confidence here can be calculated by integrating multiple verification parameters, such as the degree of content matching between structured invoice information and reimbursement reason information, the compliance verification results of structured invoice information and reimbursement reason information, the sensitivity risk of reimbursement reason information and / or the logical rationality of reimbursement reason information in real-time business scenarios.

[0092] Step 150: Based on the decision component, and combining the identification confidence, the verification confidence, and the cognitive confidence, a decision is made to obtain the review result of the document to be reviewed.

[0093] like Figure 4 As shown, the decision component (also known as the decision layer or decision fusion layer) aims to organically integrate the results of hard rule verification with the results of soft semantic reasoning. The core components of this decision layer include a decision fusion center, a weighted fusion algorithm, a conflict resolution mechanism, a confidence aggregation module, and a decision output module.

[0094] As the control center of the decision-making layer, the decision fusion center is responsible for receiving recognition confidence from the perception layer, verification confidence from the verification layer, and cognitive confidence from the cognition layer.

[0095] Optionally, after obtaining the confidence levels of the above three dimensions, the following steps can be scheduled and executed through the decision fusion center to achieve comprehensive decision-making: First, a weighted fusion algorithm is invoked to configure the weights. Specifically, this algorithm dynamically allocates weights based on the reliability of information from different sources. For example, considering the potential for errors in OCR recognition, its weight in the overall score calculation can be appropriately reduced when the recognition confidence is low; while given the complexity of cognitive reasoning, specific weights can be assigned to cognitive confidence.

[0096] Subsequently, the weights output by the weighted fusion algorithm are used as the final weights; alternatively, a conflict resolution mechanism is invoked to handle discrepancies in conclusions between levels. If there is a contradiction in the output conclusions between any levels, the weights corresponding to the confidence levels output by each component are corrected or adjudicated, and then used as the final weights. For example, if the verification layer determines compliance but the cognition layer determines doubt, the weights corresponding to the confidence levels output by each component are directly corrected or adjudicated according to a preset priority strategy or negotiation logic to obtain the final weights.

[0097] Subsequently, the confidence aggregation module, based on the final weights, weights the confidence scores of each component to obtain a comprehensive confidence score reflecting the overall credibility. The decision output module then uses this calculated comprehensive confidence score to determine the review result for the document to be reviewed. This represents a leap from single-source judgment to multi-source collaborative decision-making, significantly improving the robustness and accuracy of document review in complex and ambiguous scenarios. The review result typically includes classifications such as approved, manually reviewed, or rejected.

[0098] For example, in one possible implementation, the review result of the document to be reviewed is determined by combining the identification confidence, the verification confidence, and the cognitive confidence, including: The recognition confidence, verification confidence, and cognition confidence are weighted and fused to obtain a comprehensive confidence score. If the overall confidence level is greater than or equal to the first threshold, the review result is determined to be a successful review result. If the overall confidence level is less than the first threshold and greater than the second threshold, the review result is determined to be the result that triggers manual review. If the overall confidence level is less than the second threshold, the review result is determined to be a rejection.

[0099] like Figure 4 As shown, the final weights of each confidence level are first called in the confidence aggregation module, and the confidence levels are weighted and fused to obtain the comprehensive confidence level.

[0100] Then, the decision output module is invoked to execute the following dual-threshold strategy for classification: If the overall confidence level is greater than or equal to the first threshold (e.g., 90%), the document to be reviewed is determined to be highly credible and compliant, and automated review is triggered directly without human intervention.

[0101] If the overall confidence level is less than the first threshold but greater than the second threshold (e.g., 60%), and it is determined that the audit result is suspicious and uncertain whether it is a genuine violation, the audit result can be determined as a result that triggers manual review. This will push the audit task to the manual review interface, trigger the manual review process, and highlight key inspection items, such as risk factors that lead to a decrease in confidence level.

[0102] If the overall confidence level is less than the second threshold (e.g., 60%), it is considered that there is a clear violation (such as invoice fraud) or an extremely high risk. The review result can be determined as a rejection, so as to directly block the application and provide the user with specific reasons for rejection and modification suggestions.

[0103] The method provided in this embodiment constructs a decision fusion mechanism that integrates hard rules and soft reasoning through decision components, and uses a dual-threshold grading strategy to perform stratified judgment on the comprehensive confidence level obtained by weighted calculation. This not only ensures the automated processing efficiency of high-confidence samples, but also effectively controls the compliance risk of low-confidence samples through targeted risk interception and manual review, thereby achieving a dual improvement in review efficiency and risk control accuracy.

[0104] Furthermore, after obtaining the review results of the documents to be reviewed, the output layer can display these results and generate detailed review reports. These reports record the analysis process and evidence chain at each level, as well as a complete report of risk points and recommendations, supporting audit traceability and system optimization. Additionally, it supports feeding back the results of manual review to the experience replay pool for continuous model iteration, and can also update the rule base based on the results of manual review, thus forming a closed-loop iterative capability.

[0105] The method provided in this embodiment dynamically generates a target task chain based on the indication information in the target input information, breaking the rigidity of the traditional hard-coded process and realizing the flexible reconstruction of the review logic to effectively adapt to dynamic business changes. At the same time, by combining the perception component, verification component, and cognition component in the task chain to conduct multi-dimensional in-depth analysis, it fills the gap in semantic understanding of traditional keyword matching. Finally, the decision component combines the recognition confidence, verification confidence, and cognition confidence outputs of the multi-component decision-making to make a comprehensive judgment, effectively avoiding the one-sidedness of single logical judgment. Thus, while significantly improving the flexibility of the review process, it greatly enhances the reliability of identifying hidden risks in complex reimbursement scenarios.

[0106] In some embodiments, step 140 specifically includes: The process of performing consistency and compliance reasoning on the structured invoice information and the reimbursement reason information in the document to be audited to obtain the cognitive confidence level of the document to be audited includes: Based on the invoice type and the industry type of the seller in the structured invoice information, obtain multi-source information associated with the document to be reviewed; the multi-source information includes enterprise qualification data, historical review case records, compliance strategy knowledge base, and sensitive word database; Based on the identification confidence and the verification confidence, target constraint information is generated; The structured invoice information, the reimbursement reason information, the multi-source information, and the target constraint information are input into the inference model to obtain the cognitive confidence.

[0107] Figure 5 This is a schematic diagram of the cognitive confidence acquisition process provided by the present invention; for example... Figure 5 As shown, the cognitive component can obtain cognitive confidence through the following steps: First, multimodal information is obtained through a multi-source information input layer. The specific implementation steps are as follows: The system obtains structured invoice information and identification confidence from the output of the perception component, verification confidence from the output of the verification component, and reimbursement reason information from the documents to be reviewed. This reimbursement reason information is a natural language description filled in by the user in the reimbursement system, which expresses the reason for reimbursement and its background.

[0108] Simultaneously, the system parses the classification information from the structured invoice information to obtain the invoice type and the industry type of the seller. Based on the invoice type and the seller's industry type, it automatically retrieves related multi-source information. This multi-source information includes, but is not limited to, enterprise qualification data, historical audit case records, compliance strategy knowledge base, and sensitive word database. The enterprise qualification data is external compliance data such as the seller's business scope, enterprise status, and credit rating, obtained by calling external interfaces, used to assist in judging the authenticity of the transaction background. Historical audit case records are obtained by retrieving past audit results and processing opinions for similar scenarios from the database. The compliance strategy knowledge base is formed by loading the specific reimbursement standards, prohibited items list, and approval rules formulated internally by the enterprise. The sensitive word database is a database that loads keywords containing violations, used to quickly screen high-risk expressions.

[0109] Furthermore, the recognition confidence level output from the perception layer and the verification confidence level output from the verification layer are transformed into target constraint information. This target constraint information is then injected into the inference process as a structural context constraint. This target constraint information plays two specific roles: the recognition confidence level serves as a data credibility reference, adjusting the inference model's reliance on structured invoice information. For example, when the recognition confidence level of a certain field is low, this constraint information instructs the inference model to reduce its reliance on that field's value during inference, instead relying more on the user-entered reimbursement reason text and external API verification information for cross-judgment, thereby improving the robustness of the inference conclusion. The verification confidence level provides the inference model with verified objective fact constraints. For instance, when the verification layer confirms that "the seller's company status is cancelled," the inference model will incorporate this conclusion as a premise into the inference chain when analyzing the rationality of the reimbursement reason, ensuring that semantic inference does not contradict verified objective facts.

[0110] After obtaining multimodal information, namely structured invoice information, reimbursement reason information, multi-source information, and target constraint information, through the multi-source information input layer, the multimodal information can be fed into the inference model. The inference model performs semantic analysis on the multimodal information and outputs a cognitive confidence level that reflects the compliance possibility of the reimbursement behavior. This achieves deep integration and deep semantic understanding of various heterogeneous information in the reimbursement application, improving the accuracy, compliance, and intelligence level of reimbursement review.

[0111] The inference model can be an existing pre-trained language model trained on a large number of text pairs, or a pre-trained language model trained on a large number of text pairs with a special structure configuration, such as a language model built with a multimodal encoder and various attention networks. This embodiment does not specifically limit it.

[0112] It should be noted that the inference model here can be trained by performing the following training steps in the model training and optimization layers: First, a lightweight fine-tuning of the pre-trained language model is performed using the large model fine-tuning module and the Low-Rank Adaptation (LoRA) technique. Specifically, while keeping the main parameters of the pre-trained model frozen, trainable low-rank decomposition matrices are injected into specific layers, and the rank is set, for example, to 8. By loading an instruction fine-tuning dataset constructed for financial auditing scenarios—that is, a dataset containing sample auditing documents, multimodal information of sample auditing documents, and auditing labels of sample auditing documents—only a few adapter parameters need to be updated, allowing the model to quickly adapt to the complex logical reasoning in specific auditing tasks. This method significantly reduces computational costs and the risk of overfitting, achieving efficient model adaptation.

[0113] Next, to further improve the model's performance in real-world business scenarios, reinforcement learning optimization was performed on the finely tuned model using the experience replay pool and the PPO collaborative optimization module. Specifically, the experience replay pool continuously collects historical review records and their corresponding final review feedback from the historical review process, i.e., the review results after manual review, forming a high-quality training dataset.

[0114] Based on this dataset, the Proximal Policy Optimization (PPO) algorithm is used to continuously optimize the behavior strategies of the fine-tuned model and the collaborative OCR model through reinforcement learning, so as to obtain a real-time updated inference model.

[0115] In this process, a specific reward function R can be introduced to drive the model update, as defined below: R=0.7 Rbase-100 Rthreat+0.3 Rcognitive; Rbase is the base accuracy reward; Rthreat is the risk penalty for missing or misidentifying key risk fields; and Rcognitive is the cognitive rationality reward.

[0116] Therefore, by maximizing this reward function, the behavioral strategies of OCR and large language models can be jointly optimized, thereby effectively balancing basic accuracy, risk avoidance, and cognitive rationality, and thus obtaining a real-time updated inference model.

[0117] The method provided in this embodiment constructs a complete chain of evidence by introducing multi-source information and target constraint information based on identification and verification confidence. This not only enables the inference model to make judgments based on sufficient business context, but more importantly, it effectively avoids contradictions between model inference and objective verification facts through structured context constraints, and automatically adjusts dependency weights when data quality is poor, significantly improving the accuracy and robustness of the inference results.

[0118] In some embodiments, the step of inputting the structured invoice information, the reimbursement reason information, the multi-source information, and the target constraint information into the inference model to obtain the cognitive confidence includes: The structured invoice information, the reimbursement reason information, the multi-source information, and the target constraint information are input into the feature encoding module corresponding to the data type of each piece of information in the inference model to obtain the feature vector of each piece of information. The feature vectors of each piece of information are input into the attention module in the inference model to obtain the semantic similarity between the feature vectors of each piece of information; Based on the semantic similarity, the feature vectors of each piece of information are weighted and fused to obtain the fused feature. The fused features are input into the reasoning module of the reasoning model to obtain the cognitive confidence.

[0119] like Figure 5 As shown, the inference model includes at least a feature encoding and vectorization module, an attention module (also known as a cross-modal attention fusion layer), an inference module, and an output module.

[0120] The feature encoding and vectorization module is configured with different feature encoding modules for different data types. The attention module is configured with at least one type of attention mechanism unit, such as at least one of self-attention units, cross-attention units, and multi-head attention units. The inference module is configured with at least one inference submodule to perform audit inference in at least one dimension.

[0121] Furthermore, after obtaining multimodal information, namely structured invoice information, reimbursement reason information, multi-source information, and target constraint information through the multi-source information input layer, the multimodal information can be first input into the feature encoding and vectorization layer in the inference model. The feature encoding module corresponding to the data type of each piece of information configured in the feature encoding and vectorization layer will encode the corresponding features of each piece of information to obtain the feature vector of each piece of information, thereby realizing the unified representation of cross-modal information.

[0122] For example, for natural language text (such as information about the reason for reimbursement), a bidirectional encoder representation from transformers (BERT) is used to process the text sequence into a high-dimensional semantic vector to capture its contextual semantic relationships. For structured data (such as enterprise qualification information, historical audit case records, structured invoice information, etc.), a graph neural network (GNN) is used for processing. By modeling the complex dependency structure between entities, node embedding vectors are generated.

[0123] Finally, the heterogeneous vectors output by different feature encoding modules are mapped to the same semantic space through the feature embedding layer, ensuring that different types of information can be compared and fused in the same dimension to obtain the feature vectors of each information. This achieves a standardized representation of multi-source heterogeneous information and provides a consistent input basis for the subsequent attention mechanism.

[0124] Furthermore, after obtaining the feature vectors of each piece of information, in order to effectively integrate information from different sources and dynamically adjust their weights, the feature vectors of each piece of information can be input into the attention module, so that the attention module can realize the interaction between information, obtain the semantic similarity between the feature vectors of each piece of information, and determine the weight of each feature vector based on the semantic similarity. Then, the feature vectors of each piece of information are weighted and fused to obtain fused features. In this way, the decision-making flexibility and reliability can be improved by autonomously learning the importance of information.

[0125] For example, after obtaining the semantic similarity between each feature vector, the corresponding weights can be calculated using the Softmax function to automatically focus on the most discriminative information source, significantly improving the robustness and interpretability of the reasoning process. The specific calculation formula is as follows: α_i = exp(score_i) / Σexp(score_j); Wherein, score_j and score_i are the importance scores of feature vector j and feature vector i calculated based on the semantic similarity between each feature vector; α_i is the weight of feature vector i.

[0126] After obtaining the fusion features, the fusion features can be input into the inference module. Based on the fusion features, the inference model can perform consistency and compliance inference on the structured invoice information and the reimbursement reason information in the documents to be reviewed. Finally, the output module quantifies the credibility of the model prediction, i.e., cognitive confidence.

[0127] The method provided in this embodiment achieves deep semantic alignment and fusion of cross-modal information through multi-source heterogeneous feature encoding, attention mechanism weighting, and Softmax dynamic weight allocation. It effectively solves the risk control blind spot problem caused by isolated processing of multi-source information in traditional solutions, and improves the accuracy and robustness of identifying hidden risks in complex reimbursement scenarios.

[0128] In some embodiments, inputting the feature vectors of each piece of information into the attention module of the inference model to obtain the semantic similarity between the feature vectors of each piece of information includes: The feature vectors of each piece of information are input into the self-attention unit in the attention module to obtain the local semantic enhancement features of each piece of information. The feature vectors of each piece of information are input into the cross-attention unit in the attention module to obtain the cross-modal complementary features of each piece of information. The local semantic enhancement features and the cross-modal complementary features are input into the multi-head attention unit in the attention module to obtain the semantic similarity between the feature vectors of each piece of information.

[0129] like Figure 5As shown, the attention module includes self-attention units, cross-attention units, and multi-head attention units; correspondingly, the attention module can obtain the semantic similarity between the feature vectors of each piece of information through the following steps: First, the feature vectors of each piece of information are input into the self-attention unit in the attention module. This unit is responsible for modeling the internal correlations of information within a single modality, that is, calculating the dependencies between different elements within the same modality, so as to effectively enhance the local semantic expressive power of a single modality and obtain local semantically enhanced features.

[0130] Simultaneously, the feature vectors of each piece of information are input into the cross-modal attention unit within the attention module. This unit is responsible for achieving information complementarity and alignment between different modalities. Specifically, this unit calculates the correlation matrix between the feature vectors of one modality and the feature vectors of another modality. For example, it performs semantic matching between the user's description of the reimbursement reason and the structured invoice information extracted by OCR, in order to obtain cross-modal complementary features of each piece of information.

[0131] Through the above cross-calculation, the deep logical connections between modalities can be captured, thereby obtaining cross-modal complementary features.

[0132] Finally, the local semantic enhancement features and cross-modal complementary features are input into the multi-head attention unit in the attention module, such as an 8-head attention unit. This unit is responsible for computing the attention distribution in parallel across multiple independent semantic subspaces to capture the comprehensive semantic similarity between the feature vectors of each piece of information.

[0133] The method provided in this embodiment, through a multi-layered processing architecture of self-attention, cross-attention, and multi-head attention, not only understands the internal meaning of single data, but also deeply understands the deep logical connections between image, text, and knowledge base data, significantly enhancing the ability to identify hidden risks such as discrepancies between images and text or logical contradictions.

[0134] In some embodiments, inputting the fused features into the inference module of the inference model to obtain the cognitive confidence includes: The fused features are input into the multi-task reasoning unit in the reasoning module to obtain multi-dimensional reasoning results; the multi-dimensional reasoning results include multiple items from consistency verification results, compliance verification results, sensitive content identification results, and contextual reasoning results; The multi-dimensional reasoning results are input into the result aggregation unit in the reasoning module to obtain the cognitive confidence. The consistency verification result includes the degree of content matching between the structured invoice information and the reimbursement reason information, and the degree of qualification matching between the structured invoice information and the enterprise qualification data; the compliance verification result includes the degree of conformity between the reimbursement reason information and each preset compliance strategy data in the compliance strategy knowledge base; the sensitive content identification result includes the degree of matching between the reimbursement reason information and each preset sensitive word in the sensitive word library; the context reasoning result includes the logical reasonableness score of the reimbursement reason information in the current business context obtained by analogy reasoning through the historical audit case records.

[0135] like Figure 5 As shown, the reasoning module can generate cognitive confidence through the following steps: First, the fused features of multimodal information are input into the multi-task reasoning unit in the reasoning module. This multi-task reasoning unit contains four core reasoning sub-modules: information consistency verification sub-module, semantic compliance analysis sub-module, sensitive content identification sub-module, and context understanding sub-module. It supports progressive logical deduction under complex conditions, i.e., multi-hop reasoning, thereby outputting multi-dimensional reasoning results in parallel.

[0136] Specifically, the information consistency verification submodule can compare the structured invoice information (such as the amount and time extracted by OCR) with the reimbursement reason information (such as the amount and time description declared by the user) based on the fusion characteristics to see if they are consistent. It can also compare the enterprise information in the structured invoice information with the enterprise qualification data to obtain the consistency verification result (also known as the consistency result), thereby reflecting whether the logic between the data is consistent.

[0137] The semantic compliance analysis submodule utilizes fusion features to determine the degree of conformity between the expense reimbursement information and the data of each preset compliance strategy in the compliance strategy knowledge base. For example, it combines the company's travel standards to determine whether the reason for business banquets complies with the current reimbursement limit, so as to obtain the compliance verification result (also known as the compliance result), and then determine whether the expense reimbursement information meets the requirements of the company's system.

[0138] The sensitive content identification submodule uses fusion features to determine the degree of matching between the reimbursement reason information and each preset sensitive word in the sensitive word library. For example, it detects whether there are potential violations in the reimbursement reason information to obtain the sensitive content identification result (also known as the risk level).

[0139] The context understanding submodule utilizes fusion features to assess the logical reasonableness score of reimbursement reasons obtained through analogy reasoning from historical review case records in the current business context. For example, by combining similar historical cases with the current context, it evaluates the rationality and abnormality of this reimbursement behavior in terms of business logic and obtains a logical reasonableness score (also known as logical reasonableness).

[0140] Subsequently, the multi-dimensional reasoning results generated above are quantified and weighted, or logistic regression is performed on the multi-dimensional reasoning results generated above to obtain a score that comprehensively considers multiple dimensions such as consistency, compliance, sensitivity and rationality. Finally, the credibility of the reasoning model prediction is quantified and output, namely cognitive confidence.

[0141] The method provided in this embodiment, through the multi-hop, interpretable deep reasoning capabilities of the multi-task reasoning unit, performs in-depth analysis of consistency, compliance, sensitive content, and contextual logic in parallel, and weights and aggregates the multi-dimensional reasoning results. This not only achieves a comprehensive perspective on reimbursement behavior from data matching to semantic rationality, but also makes the generated cognitive confidence level have fine-grained interpretability, effectively breaking through the limitations of traditional single-rule verification, and significantly improving the accuracy and anti-interference capability of intelligent review in complex scenarios.

[0142] The method provided in this embodiment will be described in detail below with specific examples.

[0143] Figure 6 This is the second flowchart illustrating the document review method provided by this invention; for example... Figure 6 As shown, when a target task chain is constructed based on the document review intent corresponding to the target input information, and the target task chain includes a perception component, a verification component, a cognition component, and a decision-making component, the following end-to-end automated review is achieved through multimodal information fusion and multi-agent collaborative reasoning, with the dynamically reconfigurable target task chain as the core driving mechanism: First, receive the image of the invoice to be reviewed and the reason for the reimbursement from the user; Secondly, the uploaded files undergo preliminary verification, including but not limited to file integrity checks, format validity verification, and image quality assessments such as sharpness, tilt, and lighting uniformity. If a file does not meet the processing requirements, an error message will be provided immediately, guiding the user to re-upload.

[0144] If the file meets the processing requirements, the perception component first uses the OCR model to identify the key fields of the invoice image to be reviewed and classify the invoice. The key field content recognition and invoice classification are then converted into standardized structured data objects, clarifying the semantic role of each field to obtain structured invoice information, and attaching the corresponding recognition confidence level.

[0145] Then, by verifying the component, the corresponding compliance strategy knowledge base and third-party data are dynamically loaded using the invoice type in the structured invoice information. The compliance of the structured invoice information is verified through the compliance strategy knowledge base and third-party data. The number of duplicate reimbursements and the frequency of anomalies in the structured invoice information are verified through the historical reimbursement record database. Finally, the verification results of these two aspects are quantified to obtain the verification confidence of the structured invoice information.

[0146] Then, through the cognitive component, the system utilizes the invoice type and the industry type of the seller in the structured invoice information to obtain enterprise qualification data, historical audit case records, compliance strategy knowledge base, and sensitive word database. The system then uses an inference model to perform multimodal fusion decision-making on the enterprise qualification data, historical audit case records, compliance strategy knowledge base, sensitive word database, structured invoice information, reimbursement reason information, and target constraint information including the identification confidence and verification confidence of the upper layer output, in order to obtain the cognitive confidence.

[0147] Then, through the decision-making component, the confidence levels output by each component in the perception, verification, and cognition components are integrated using a dynamic weighting strategy to generate a comprehensive confidence level. A threshold judgment is then applied based on this comprehensive confidence level. If the comprehensive confidence level is greater than or equal to the first threshold (e.g., 90%), the document to be reviewed is determined to be highly credible and compliant, triggering automated review directly without manual intervention. If the comprehensive confidence level is less than the first threshold but greater than the second threshold (e.g., 60%), the review result is considered questionable and uncertain regarding whether there is a genuine violation. This result triggers manual review, pushing the review task to the manual review interface and highlighting key inspection items, such as risk factors that lower the confidence level. If the comprehensive confidence level is less than the second threshold (e.g., 60%), a clear violation (e.g., invoice forgery) or extremely high risk is considered present, leading to a rejection of the application. The user is then provided with specific reasons for rejection and suggested modifications. Finally, after obtaining the audit results of the documents to be audited, the audit results can be displayed through the output layer (also known as the output component). A detailed audit report can also be generated, recording the analysis process and evidence chain at each level, as well as a complete report of risk points and suggestions, supporting audit traceability and system optimization.

[0148] Verification has shown that the method provided in this application improves efficiency by reducing the time for single-document review from 5 minutes to 5 seconds, a 59-fold increase; optimizes costs by reducing the workload of finance personnel by over 95%, freeing up high-value work; and ensures risk control by achieving an accuracy rate of over 95% in rule-based review and reducing the false positive rate to below 1.2%. Therefore, the method provided in this application yields significant technical benefits in improving efficiency, reducing costs, and strengthening risk control.

[0149] The document review device provided by the present invention is described below. The document review device described below and the document review method described above can be referred to in correspondence.

[0150] Figure 7 This is a schematic diagram of the document review device provided by the present invention; as shown below. Figure 7 As shown, the device includes: The generation unit 710 is used to generate a target task chain based on the document review intent corresponding to the target input information; the target input information includes instruction information for configuring the task chain; The perception unit 720 is used to extract structured information from the image of the invoice to be reviewed in the document to be reviewed based on the perception component when the functional components of the target task chain include a perception component, a verification component, a cognition component and a decision component, thereby obtaining structured invoice information and the recognition confidence of the structured invoice information. The verification unit 730 is used to perform compliance verification on the structured invoice information based on the verification component, and obtain the verification confidence level of the structured invoice information; The cognitive unit 740 is used to perform consistency and compliance reasoning on the structured invoice information and the reimbursement reason information in the document to be audited based on the cognitive component, and to obtain the cognitive confidence level of the document to be audited. The decision unit 750 is used to determine the audit result of the document to be audited based on the decision component, in conjunction with the identification confidence, the verification confidence and the cognitive confidence.

[0151] The device provided in this embodiment automatically parses task objectives and logical conditions based on the document review intent, and dynamically retrieves components, configures parameters, and plans execution paths accordingly. This achieves end-to-end automated mapping from abstract document review intent to specific task chains. This mechanism deeply decouples the review business logic from the underlying code, completely eliminating the dependence on fixed templates or hard-coded development. Users can reconstruct the review process in real time simply by providing instructions. In other words, users only need to input commands to automatically complete intent recognition, component matching, connection inference, and parameter setting, completely eliminating the dependence on code development. This avoids the problem that traditional low-code platforms still require manual intervention when facing complex business changes. Even users without a programming background can dynamically adjust and deploy complex review processes through simple text commands to quickly respond to business changes and achieve real-time process updates. This not only significantly reduces operation and maintenance costs and shortens the response cycle from the traditional weekly level to the real-time level, but also supports the dynamic evolution and intelligent completion of task chains, achieving end-to-end automated mapping from abstract business intent to specific task chains, thus improving the flexibility and reliability of the review process.

[0152] The apparatus provided by the present invention is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0153] like Figure 2 As shown, this application also provides a document review system, which includes a user interaction module, a task chain engine, a processor, a database, and a model optimizer that communicate with each other. The task chain engine includes a memory, an execution engine, and a computer program stored in the memory and running on the processor. The execution engine can call logical instructions in the memory to execute a document review method. This method includes: generating a target task chain based on the document review intent corresponding to target input information; the target input information includes instruction information for configuring the task chain; when the functional components of the target task chain include a perception component, a verification component, a cognition component, and a decision component, based on the perception component, performing structured information extraction on the image of the invoice to be reviewed in the document to be reviewed, obtaining structured invoice information and the recognition confidence level of the structured invoice information; based on the verification component, performing compliance verification on the structured invoice information, obtaining the verification confidence level of the structured invoice information; based on the cognition component, performing consistency and compliance reasoning on the structured invoice information and the reimbursement reason information in the document to be reviewed, obtaining the cognition confidence level of the document to be reviewed; based on the decision component, combining the recognition confidence level, the verification confidence level, and the cognition confidence level, deciding on the review result of the document to be reviewed.

[0154] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the executable file review method provided by the above methods. The method includes: generating a target task chain based on the file review intent corresponding to the target input information; the target input information includes instruction information for configuring the task chain; when the functional components of the target task chain include a perception component, a verification component, a cognition component, and a decision component, based on the perception component, performing structured information extraction on the image of the invoice to be reviewed in the file to be reviewed, obtaining structured invoice information and the recognition confidence of the structured invoice information; based on the verification component, performing compliance verification on the structured invoice information, obtaining the verification confidence of the structured invoice information; based on the cognition component, performing consistency and compliance reasoning on the structured invoice information and the reimbursement reason information in the file to be reviewed, obtaining the cognition confidence of the file to be reviewed; based on the decision component, combining the recognition confidence, the verification confidence, and the cognition confidence, deciding to obtain the review result of the file to be reviewed.

[0155] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the executable file review method provided by the above methods. This method includes: generating a target task chain based on the file review intent corresponding to target input information; the target input information including instruction information for configuring the task chain; when the functional components of the target task chain include a perception component, a verification component, a cognition component, and a decision component, based on the perception component, extracting structured information from the image of the invoice to be reviewed in the file to be reviewed, obtaining structured invoice information and the recognition confidence level of the structured invoice information; based on the verification component, performing compliance verification on the structured invoice information, obtaining the verification confidence level of the structured invoice information; based on the cognition component, performing consistency and compliance reasoning on the structured invoice information and the reimbursement reason information in the file to be reviewed, obtaining the cognition confidence level of the file to be reviewed; and based on the decision component, combining the recognition confidence level, the verification confidence level, and the cognition confidence level, deciding to obtain the review result of the file to be reviewed.

[0156] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0157] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A document review method, characterized in that, include: Generate a target task chain based on the document review intent corresponding to the target input information; The target input information includes instructions for configuring the task chain; When the functional components of the target task chain include a perception component, a verification component, a cognition component, and a decision-making component, based on the perception component, structured information is extracted from the image of the invoice to be reviewed in the document to be reviewed, and structured invoice information and the recognition confidence of the structured invoice information are obtained. Based on the verification component, the structured invoice information is subjected to compliance verification to obtain the verification confidence level of the structured invoice information; Based on the cognitive component, consistency and compliance reasoning are performed on the structured invoice information and the reimbursement reason information in the document to be audited to obtain the cognitive confidence level of the document to be audited. Based on the decision component, and by combining the identification confidence, the verification confidence, and the cognitive confidence, the audit result of the document to be audited is obtained.

2. The document review method according to claim 1, characterized in that, The process of performing consistency and compliance reasoning on the structured invoice information and the reimbursement reason information in the document to be audited to obtain the cognitive confidence level of the document to be audited includes: Based on the invoice type and the industry type of the seller in the structured invoice information, obtain multi-source information associated with the document to be reviewed; the multi-source information includes enterprise qualification data, historical review case records, compliance strategy knowledge base, and sensitive word database; Based on the identification confidence and the verification confidence, target constraint information is generated; The structured invoice information, the reimbursement reason information, the multi-source information, and the target constraint information are input into the inference model to obtain the cognitive confidence.

3. The document review method according to claim 2, characterized in that, The step of inputting the structured invoice information, the reimbursement reason information, the multi-source information, and the target constraint information into the inference model to obtain the cognitive confidence includes: The structured invoice information, the reimbursement reason information, the multi-source information, and the target constraint information are input into the feature encoding module corresponding to the data type of each piece of information in the inference model to obtain the feature vector of each piece of information. The feature vectors of each piece of information are input into the attention module in the inference model to obtain the semantic similarity between the feature vectors of each piece of information; Based on the semantic similarity, the feature vectors of each piece of information are weighted and fused to obtain the fused feature. The fused features are input into the reasoning module of the reasoning model to obtain the cognitive confidence.

4. The document review method according to claim 3, characterized in that, The step of inputting the feature vectors of each piece of information into the attention module of the inference model to obtain the semantic similarity between the feature vectors of each piece of information includes: The feature vectors of each piece of information are input into the self-attention unit in the attention module to obtain the local semantic enhancement features of each piece of information. The feature vectors of each piece of information are input into the cross-attention unit in the attention module to obtain the cross-modal complementary features of each piece of information. The local semantic enhancement features and the cross-modal complementary features are input into the multi-head attention unit in the attention module to obtain the semantic similarity between the feature vectors of each piece of information.

5. The document review method according to claim 3, characterized in that, The step of inputting the fused features into the inference module of the inference model to obtain the cognitive confidence includes: The fused features are input into the multi-task reasoning unit in the reasoning module to obtain multi-dimensional reasoning results; the multi-dimensional reasoning results include multiple items from consistency verification results, compliance verification results, sensitive content identification results, and contextual reasoning results; The multi-dimensional reasoning results are input into the result aggregation unit in the reasoning module to obtain the cognitive confidence. The consistency verification result includes the degree of content matching between the structured invoice information and the reimbursement reason information, and the degree of qualification matching between the structured invoice information and the enterprise qualification data; the compliance verification result includes the degree of conformity between the reimbursement reason information and each preset compliance strategy data in the compliance strategy knowledge base; the sensitive content identification result includes the degree of matching between the reimbursement reason information and each preset sensitive word in the sensitive word library; the context reasoning result includes the logical reasonableness score of the reimbursement reason information in the current business context obtained by analogy reasoning through the historical audit case records.

6. The document review method according to any one of claims 1-5, characterized in that, The compliance verification of the structured invoice information, to obtain the verification confidence level of the structured invoice information, includes: Based on the bill type in the structured bill information, determine the compliance strategy knowledge base and third-party data; Based on the compliance strategy knowledge base and the third-party data, the compliance of the structured invoice information is verified to obtain a first verification result; Based on the historical reimbursement record database, the number of duplicate reimbursements and the frequency of abnormalities in the structured invoice information are verified to obtain a second verification result; Based on a preset quantization mapping strategy, the first verification result and the second verification result are comprehensively scored to obtain the verification confidence level.

7. The document review method according to any one of claims 1-5, characterized in that, The step of generating a target task chain based on the document review intent corresponding to the target input information includes: Based on the document review intent, obtain the task objective, target component type, and logical control conditions; Based on the target component type, retrieve multiple functional components from the prefabricated component library; Based on the stated task objectives, determine the operating parameters of each of the stated functional components; Based on the aforementioned logical control conditions, the data flow paths and execution dependencies between the functional components are determined. Based on the data flow path, the execution dependencies, and the running parameters, multiple functional components are encapsulated to generate the target task chain.

8. The document review method according to any one of claims 1-5, characterized in that, The decision to obtain the review result of the document to be reviewed by combining the identification confidence, the verification confidence, and the cognitive confidence includes: The recognition confidence, verification confidence, and cognition confidence are weighted and fused to obtain a comprehensive confidence score. If the overall confidence level is greater than or equal to the first threshold, the review result is determined to be a successful review result. If the overall confidence level is less than the first threshold and greater than the second threshold, the review result is determined to be the result that triggers manual review. If the overall confidence level is less than the second threshold, the review result is determined to be a rejection.

9. A document review device, characterized in that, include: The generation unit is used to generate a target task chain based on the document review intent corresponding to the target input information; The target input information includes instructions for configuring the task chain; The perception unit is used to extract structured information from the image of the invoice to be reviewed in the document to be reviewed based on the perception component, when the functional components of the target task chain include a perception component, a verification component, a cognition component and a decision component, to obtain structured invoice information and the recognition confidence of the structured invoice information. The verification unit is used to perform compliance verification on the structured invoice information based on the verification component, and obtain the verification confidence level of the structured invoice information; A cognitive unit is used to perform consistency and compliance reasoning on the structured invoice information and the reimbursement reason information in the document to be audited based on the cognitive component, and to obtain the cognitive confidence level of the document to be audited. The decision unit is used to determine the audit result of the document to be audited based on the decision component, in conjunction with the identification confidence, the verification confidence, and the cognitive confidence.

10. A document review system, characterized in that, This includes a user interaction module, a task chain engine, a processor, a database, and a model optimizer; The task chain engine includes a memory, an execution engine, and a computer program stored in the memory and running on the processor. When the execution engine executes the computer program, it implements the document review method as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the document review method as described in any one of claims 1 to 8.

Images